Law enforcement recorder voice intercom system with fusion intelligent sensor

Abstract

The present application relates to the technical field of intelligent sensor, in particular to a law enforcement recorder voice intercom system fusing intelligent sensor, which comprises a voice frame processor, a sensing state fusion processor, a situation classification processor and an intercom control processor, the voice frame processor extracts short-time energy, harmonic structure, speech intelligibility, noise spectrum stability and speech endpoint candidate probability, the sensing state fusion processor establishes a sensing situation vector under a unified time reference by combining motion intensity, posture change rate, wearing state stability, environmental sound pressure fluctuation and device interaction state with the voice frame, the situation classification processor generates a speech credible mask and a link protection level accordingly, and the intercom control processor jointly configures a noise suppression weight, a coding code rate, a forward error correction proportion and a sending buffer control parameter. The system can distinguish near-talking voice, frictional obstruction, short-time impact and strong noise interference, and improve speech continuity and link protection matching in a mobile law enforcement scene.

Description

Technical Field

[0001] This invention relates to the field of intelligent sensor technology, specifically to a voice intercom system for law enforcement recorders that integrates intelligent sensors. Background Technology

[0002] Law enforcement recorder voice intercom systems typically rely on the recorder's built-in microphone, voice encoding processing program, and wireless transmission link to achieve on-site voice acquisition, encoding, transmission, and playback. In conventional solutions, the voice input end first processes the audio signal acquired by the microphone into frames, then determines whether the current frame contains speech based on audio characteristics such as short-time energy, zero-crossing rate, spectral stability, and noise estimation. Segments identified as speech undergo noise reduction, echo cancellation, and compression encoding. The transmission end typically sends the encoded voice frames to the dispatch end or other law enforcement terminals according to a preset bitrate and fixed buffering strategy. The receiving end then sorts and plays them according to timestamps and sequence numbers. This type of solution primarily uses the audio itself as the basis for voice processing. Noise reduction modes, encoding parameters, and transmission methods are mostly preset at the factory or during system configuration. While it can achieve basic intercom when environmental changes are minimal, it lacks synchronous perception of the recorder's wearing status, the movement of law enforcement personnel, and changes in ambient sound pressure, making it impossible to explain the causes of audio anomalies from non-acoustic perspectives.

[0003] In more similar improvement schemes, law enforcement recorders incorporate smart sensor data as auxiliary information. For example, they use motion status to determine if the device is moving, wearing status to determine if the recorder is obstructed, and changes in ambient sound pressure to determine if the environment is noisy. These schemes typically use sensor results as scene labels, then call corresponding noise reduction or transmission modes based on the scene label. For example, they increase noise reduction intensity in high-noise conditions, increase buffer length in moving conditions, and reduce bitrate usage in stationary conditions. While this approach introduces smart sensors into the voice communication process, the correlation between sensor data and audio features is often loose. Sensor output only participates in mode switching and does not establish a joint calculation relationship with the short-time energy, harmonic structure, speech intelligibility, noise spectrum stability, and speech endpoint candidate probability of the speech frame under the same time reference. Furthermore, it does not further transform the sensor state into a speech credibility mask, noise suppression weights, and link protection levels, making it difficult for sensor data to participate in fine-grained decisions for speech enhancement and link control.

[0004] The main technical problem with existing technologies is that when law enforcement recorders are running, turning, approaching strong noise sources, experiencing clothing friction, brief obstruction, or sudden changes in device posture, the audio energy and spectral distribution of real near-field speech, wind noise, collision sounds, friction sounds, and far-field interference sounds easily overlap. Simply relying on audio features or fixed scene labels makes it difficult to distinguish the true cause of speech quality degradation, leading to a lack of unified criteria for speech denoising, endpoint detection, coding redundancy, and transmission buffer control. If the denoising intensity is too high, consonant boundaries and low-energy speech segments are easily suppressed; if the denoising intensity is insufficient, impact noise and friction noise will occupy the intercom link; if the link parameters are still configured according to fixed rules, segments containing valid speech but susceptible to interference cannot be protected accordingly. Therefore, it is necessary to establish an intercom processing mechanism that deeply integrates the intelligent sensor status with speech frame features, enabling speech enhancement and link control to coordinate and adjust based on the same sensor status. Summary of the Invention

[0005] The purpose of this invention is to provide a voice intercom system for law enforcement recorders that integrates intelligent sensors, which can solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The law enforcement recorder voice intercom system integrating intelligent sensors includes a voice frame processor, a sensor status fusion processor, a situation classification processor, and an intercom control processor. The voice frame processor is configured to frame the intercom voice collected by the law enforcement recorder and generate short-time energy, harmonic structure, speech intelligibility, noise spectrum stability and voice endpoint candidate probability. The sensor state fusion processor is configured to receive motion intensity, posture change rate, wearing status stability, ambient sound pressure fluctuation and device interaction status output by smart sensors, and establish a sensor state vector with the intercom voice frame under a unified time reference. The situation classification processor determines the intercom processing status based on the sensing situation vector and generates a voice trust mask and a link protection level. The intercom control processor configures noise suppression weights, coding rate, forward error correction ratio, and transmission buffer control parameters in conjunction with the voice trust mask and the link protection level.

[0007] Preferably, the sensor state fusion processor uses the start and end times of the intercom voice frame as the main index to resample, delay-compensate, and mark the motion intensity, the posture change rate, the wearing state stability, the ambient sound pressure fluctuation, and the device interaction state, and writes the acoustic features of each voice frame and the sensor features within the same time window into the same vector position. State difference components and sensor missing measurement indicators are retained between adjacent time windows. The sensing state vector includes acoustic state sub-vector, sensing state sub-vector, time alignment confidence, and inter-frame change sub-vector.

[0008] Preferably, the situation classification processor includes an acoustic state branch, a sensing cause branch, and a cross-modal consistency determination layer; The acoustic state branch generates candidate speech states based on the short-time energy, the harmonic structure, the speech intelligibility, and the noise spectrum stability. The sensing cause branch generates the interference cause state based on the motion intensity, the posture change rate, the wearing state stability, and the ambient sound pressure fluctuation; The cross-modal consistency determination layer jointly determines the candidate speech state and the interference cause state, and outputs one of the following intercom processing states: static near speech, moving near speech, strong noise near speech, friction blockage, short-term impact, or far-field interference.

[0009] Preferably, the situation classification processor configures voice trust mask generation rules and link protection level generation rules for each of the intercom processing states; The speech credibility mask generation rule maps the near-speech probability, sensor causation state, and inter-frame variation sub-vectors into a time-frequency mask matrix. The link protection level generation rule maps the stability of the time-frequency mask matrix, the voice endpoint candidate probability, and the time alignment confidence to one of the following transmission control categories: low-latency transmission, redundancy protection transmission, or delayed buffer transmission, and writes the transmission control category and the corresponding redundancy identifier into the frame-level control table.

[0010] Preferably, the sensor state fusion processor performs reliability encoding on the sensor situation vector; The reliability coding includes sensor sampling interval offset, consecutive missing measurement length, correlation between adjacent sensor features, and synchronization of acoustic feature abrupt changes. When there is a time offset between any sensing feature of the smart sensor and the voice frame, the sensing state fusion processor calculates the compensation weight based on the inter-frame change sub-vector, and writes the compensated sensing feature, compensation weight and reliability code together into the sensing state sub-vector, and then provides it to the situation classification processor as the input field for cross-modal consistency determination.

[0011] Preferably, the cross-modal consistency determination layer is configured with a state constraint matrix consisting of acoustic state tags, sensor cause tags, and wearing stability tags; The state constraint matrix defines the combinatorial relationship between the candidate speech state and the interference cause state; When the candidate speech state contains a high-energy mutation and the sensing cause state contains a high posture change or wearing instability, the cross-modal consistency determination layer will output candidates limited to friction occlusion or short-term impact. When the candidate speech state contains stable harmonics and the wearing of the stable tag satisfies the continuous stability condition, the output candidate will be limited to static proximity or mobile proximity.

[0012] Preferably, the intercom control processor includes an enhanced parameter generation path; The enhancement parameter generation path reads the time-frequency mask matrix, the intercom processing state, and the inter-frame change sub-vector, and generates noise spectrum update gating, impulse noise suppression coefficient, speech boundary preservation coefficient, and echo cancellation step size for each speech frame; The noise spectrum update gating determines the update open interval based on the duration of the interference cause state, the voice boundary preservation coefficient is determined based on the joint distribution of the voice endpoint candidate probability and the wearing state stability, and shares the frame number with the transmission control category.

[0013] Preferably, the intercom control processor includes a link parameter generation path; The link parameter generation path takes the transmission control category as input, and constructs a link parameter group by the coding rate, the forward error correction ratio, the number of repetitions of key voice frames and the receiver buffer depth, and sets a frame sequence number and a protection identifier for the link parameter group. For the redundant protection transmission, the link parameter generation path sets the forward error correction ratio based on the length of the continuous reliable region in the time-frequency mask matrix and the differential component of the ambient sound pressure fluctuation, and sets the number of key voice frame repetitions based on the voice endpoint candidate probability.

[0014] Preferably, the situation classification processor sets cross-frame state hysteresis constraints; The cross-frame state hysteresis constraint takes the current intercom processing state, the intercom processing state of the previous time window, the time alignment confidence, and the reliability code as inputs to generate a state transition permission flag. The state transition permission flag prohibits direct conversion from noisy proximity speech to static proximity speech when the time alignment confidence is lower than the confidence interval given by the cross-frame state hysteresis constraint or the continuous missing length exceeds the allowable length, and the speech confidence mask in adjacent time windows is smoothly synthesized according to the compensation weight.

[0015] Preferably, the intercom control processor sets up a voice frame priority queue associated with the link parameter group; The voice frame priority queue generates frame-level transmission sequence numbers according to the transmission control category, the voice endpoint candidate probability, the length of the continuous reliable region, and the differential component of the ambient sound pressure fluctuation. When the transmission buffer control parameter indicates that the buffer occupancy is close to the upper limit, the voice frame priority queue retains continuous voice segments with redundancy protection transmission categories and the corresponding key voice frame repetition counts, and performs delayed merging encoding on frames with low voice endpoint candidate probability and lack of wearing state stability support.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By combining the short-time energy, harmonic structure, speech intelligibility, noise spectrum stability, and speech endpoint candidate probability of a speech frame with the motion intensity, posture change rate, wearing stability, ambient sound pressure fluctuations, and device interaction status output by intelligent sensors to form a sensing situation vector under a unified time reference, the voice intercom system can establish a correspondence between audio anomalies and the recorder's motion, wearing, and environmental states. After situation classification processing, a speech reliability mask and link protection level are generated, so that noise suppression weights, coding bitrate, forward error correction ratio, and transmission buffer control parameters are driven by the same set of fused states. Therefore, when encountering high-energy sudden changes, the system can determine, based on posture changes and wearing stability, that the noise is closer to impact noise, friction obstruction, or near-speech, reducing the suppression of effective speech, the transmission of non-speech noise, and the invalidation of the intercom link, thus improving the continuity and intelligibility of speech transmission in complex mobile law enforcement scenarios.

[0017] 2. Furthermore, the sensor state fusion processor reliably encodes the sensor sampling interval offset, continuous missing measurement length, correlation between adjacent sensor features, and synchronization of acoustic feature abrupt changes. It then uses time alignment confidence, inter-frame change sub-vectors, and compensation weights to participate in situation classification, ensuring that short-term sensor missing measurements or sampling offsets do not directly cause processing state jumps. The situation classification processor sets cross-modal consistency judgments and cross-frame state hysteresis constraints, ensuring that the transitions between static near-field communication, moving near-field communication, noisy near-field communication, friction obstruction, short-term impact, and far-field interference have sensor basis and temporal continuity. The intercom control processor further writes the time-frequency mask matrix, transmission control category, continuous reliable region length, and ambient sound pressure difference component into the link parameter group and voice frame priority queue. This ensures that key voice segments, redundant protection segments, and low-reliability segments have different transmission orders when the buffer is close to its limit, reducing voice interruptions caused by frequent processing mode switching and buffer congestion. Attached Figure Description

[0018] Figure 1 A flowchart of the overall processing flow of a voice intercom system for a law enforcement recorder that integrates intelligent sensors; Figure 2Flowchart for sensor situation vector construction and reliability coding; Figure 3 Flowchart for situational classification, voice trust mask, and link protection level generation; Figure 4 This is a flowchart of the enhanced parameters, link parameters, and voice frame priority queue control for the intercom control processor. Detailed Implementation

[0019] In one embodiment, reference Figure 1 A voice intercom system for a law enforcement recorder integrating intelligent sensors includes a voice frame processor, a sensor state fusion processor, a situation classification processor, and an intercom control processor. The voice frame processor is used to segment the intercom voice collected by the law enforcement recorder into frames and generate short-time energy, harmonic structure, voice clarity, noise spectrum stability, and voice endpoint candidate probability. The sensor state fusion processor is used to receive motion intensity, posture change rate, wearing stability, ambient sound pressure fluctuation, and device interaction status output by the intelligent sensors, and establishes a sensor situation vector with the intercom voice frames under a unified time reference. The situation classification processor determines the intercom processing status based on the sensor situation vector and generates a voice trust mask and link protection level. The intercom control processor jointly configures noise suppression weights, coding rate, forward error correction ratio, and transmission buffer control parameters based on the voice trust mask and the link protection level. Specifically, the voice... The frame processor takes the continuous intercom voice collected by the law enforcement recorder as input, divides the continuous voice into frame-level data according to the overlapping of adjacent frames, and calculates the audio energy distribution, harmonic preservation, speech intelligibility, stability of the background noise spectrum, and candidate probability of speech endpoints within each frame. The motion intensity provided by the smart sensor is used to characterize the inertial disturbance generated by the law enforcement recorder as the wearer moves; the attitude change rate is used to characterize the degree of change in the recorder's orientation and wearing angle; the wearing state stability is used to characterize the continuity of contact between the recorder and the wearing position; the ambient sound pressure fluctuation is used to characterize the changes in the sound field at the scene; and the device interaction state is used to characterize the current state of interactive inputs such as intercom triggering, button pressing, touch screen, or voice wake-up. The sensor state fusion processor uses the time boundary of the voice frame as an index to time-align the data from the smart sensor and form a sensor state vector, which can be represented as: ; Where i represents the audio frame number. Represents the acoustic state subvector. Indicates short-term energy. Represents harmonic structure, Indicates speech clarity. Indicates noise spectrum stability. Indicates the candidate probability of the voice endpoint. Represents the sensing state sub-vector. Indicates exercise intensity. Indicates the rate of attitude change. Indicates the stability of the wearing condition. This indicates fluctuations in ambient sound pressure. Indicates the device interaction status. Indicates the confidence level of time alignment. This represents the change in sensor state between adjacent speech frames, when Take the higher value and and Maintain stability, while When the state is stable, the sensing situation vector is closer to the near-speech state. Take the higher value but , and When a sudden change occurs, the sensing situation vector is closer to the impact or strong noise state. After reading the sensing situation vector, the situation classification processor classifies the voice frame into intercom processing states such as static near-talk, moving near-talk, strong noise near-talk, friction blockage, short-term impact or far-field interference, and generates a voice credibility mask and link protection level associated with the intercom processing state. The intercom control processor simultaneously configures the noise suppression weight on the voice enhancement side and the coding rate, forward error correction ratio, and transmission buffer control parameters on the transmission side under the same frame sequence number, so that voice enhancement and link transmission no longer operate according to unrelated rules. The working principle of this embodiment is to establish a correspondence between the sound pattern reflected by the audio features and the device state reflected by the smart sensor on the same time axis, and then drive noise reduction, encoding and transmission control by the same sensing situation vector. The advantage of this embodiment is that it can provide a unified processing basis when device movement, wearing changes and environmental sound pressure fluctuations coexist, and reduce the processing confusion caused by real voice and non-voice interference when acoustic features are similar.

[0020] In one embodiment, reference Figure 2The sensor state fusion processor uses the start and end times of the intercom voice frames as the primary index to resample, delay-compensate, and assign confidence markers to the motion intensity, posture change rate, wearing stability, ambient sound pressure fluctuations, and device interaction state. It also writes the acoustic features of each voice frame and the sensor features within the same time window into the same vector location. State difference components and sensor missing measurement markers are retained between adjacent time windows. The sensor state vector includes an acoustic state sub-vector, a sensor state sub-vector, time alignment confidence, and an inter-frame change sub-vector. Specifically, the sensor state fusion processor processes the received voice frames... After outputting the frame-level acoustic features, the start and end times of each speech frame are read, and the corresponding smart sensor sampling points are searched between the start and end times. When the sensor sampling points are not consistent with the speech frame boundaries, linear interpolation, neighborhood preservation, or weighted smoothing is used to resample the sensor sampling results into frame-level sensing features. For time offsets caused by processing link buffering, sensor data reporting delays, or speech framing processing, the sensor state fusion processor calculates the offset between the speech frame center time and the effective sensor sampling time, and writes the offset into the confidence flag field. The time alignment confidence can be expressed as: ; in, This represents the time offset between the center time of the i-th frame of speech and the center time of the matched sensor state. This represents the time scale used for normalizing the time offset. This indicates a sensor measurement failure flag with a value of 0 or 1. Indicates the correlation between adjacent sensing features, when Approaching 0 Take 0 and When it approaches 1, Approaching a high confidence state, sensor features can directly participate in situation classification. Larger or When taking 1, The weight of the sensing branch is reduced and weakened in subsequent classifications; the inter-frame change sub-vector is obtained by differentiating the sensing states of adjacent time windows. After differentiating motion intensity, posture change rate, wearing stability and ambient sound pressure fluctuation, the direction and amplitude are retained, enabling the system to distinguish stable movement, sudden turning, short-term occlusion and sudden sound pressure changes; the sensor missing detection flag is used to record the state in which the smart sensor data has not arrived completely in a certain frame, avoiding misjudging missing detection as a stable state; the advantage of this embodiment is that through the master index, resampling, delay compensation and confidence flags, calculable frame-level fusion data is formed, so that the sensor output no longer participates in the processing with coarse-grained scene labels, but enters the subsequent classification and control logic as a vector field corresponding to each speech frame, reducing the interference caused by sampling asynchrony to speech enhancement and link control.

[0021] In one embodiment, reference Figure 3 The situation classification processor includes an acoustic state branch, a sensing cause branch, and a cross-modal consistency determination layer. The acoustic state branch generates candidate speech states based on the short-time energy, the harmonic structure, the speech intelligibility, and the noise spectrum stability. The sensing cause branch generates interference cause states based on the motion intensity, the rate of posture change, the wearing stability, and the ambient sound pressure fluctuation. The cross-modal consistency determination layer jointly determines the candidate speech states and the interference cause states, and outputs one of the following intercom processing options: static proximity intercom, moving proximity intercom, strong noise proximity intercom, friction blockage, short-time impact, or far-field interference. Specifically, the acoustic state branch normalizes the acoustic state sub-vectors and combines short-time energy and noise spectrum stability into an energy anomaly field, harmonic structure and speech intelligibility into a speech fidelity field, and speech endpoint candidate probability as a time boundary field; the sensing cause branch applies differential constraints to motion intensity, posture change rate, wearing state stability, and environmental sound pressure fluctuations, forming motion disturbance, posture change, wearing occlusion, and sound pressure change fields; the cross-modal consistency determination layer does not directly provide a classification based on a single acoustic feature, but instead calculates a joint score for different processing states, which can be expressed as: ; in, Let i represent the score vector corresponding to each intercom processing status in the i-th frame. Represents the acoustic state weight matrix. Represents the sensor state weight matrix. This represents the cross-modal interaction weight matrix. This represents the weight matrix indicating the continuity of the state in the previous frame. Represents the rating vector of the previous frame; when The short-time energy and harmonic structure in the middle simultaneously supports the presence of speech and When the stability of the wearing state is taken as the stable state, The corresponding proximity interaction item increases, making the static or mobile proximity score dominant; when the short-term energy increases but the posture change rate and the ambient sound pressure fluctuation change synchronously, the short-term impact or strong noise proximity score increases; when the speech clarity decreases and the wearing stability decreases, the friction blockage score increases; the advantage of this embodiment is that by constraining the processing state through acoustic state, sensing cause and cross-modal consistency, high-energy noise, real proximity speech and wearing friction can still be placed in different processing paths when the acoustic form is similar, reducing the misfit of fixed noise reduction mode to complex field conditions.

[0022] In one embodiment, the situation classification processor configures a voice reliability mask generation rule and a link protection level generation rule for each intercom processing state. The voice reliability mask generation rule maps the near-speech probability, sensor cause state, and inter-frame variation sub-vector into a time-frequency mask matrix. The link protection level generation rule maps the stability of the time-frequency mask matrix, the voice endpoint candidate probability, and the time alignment confidence into one of the following transmission control categories: low-latency transmission, redundancy protection transmission, or delayed buffer transmission. The transmission control category and its corresponding redundancy identifier are then written into a frame-level control table. Specifically, after outputting the intercom processing state, the situation classification processor establishes a time-frequency unit index for each voice frame and generates a time-frequency mask matrix based on the near-speech probability, impulse noise probability, friction obstruction probability, far-field interference probability, and inter-frame variation sub-vector. Each element in the time-frequency mask matrix represents the degree of voice preservation at the corresponding time and frequency band position. The time-frequency mask matrix can be represented as: ; in, This represents the mask value for the j-th frequency band in the i-th frame. This means limiting the calculation result to between 0 and 1. to This represents the mapping coefficient for each input item. Indicates the probability of near-speech speech. Indicates the probability of a short-term shock. Indicates the probability of friction occlusion. Indicates the candidate probability of the voice endpoint. Indicates the stability of the wearing condition. This represents the inter-frame variation of ambient sound pressure fluctuations; when , and When taking a higher state, The tendency is to preserve speech components, when , or When it rises, The system tends to suppress unstable components. The link protection level generation rule is based on the stability of the time-frequency mask matrix in adjacent frames, combined with the voice endpoint candidate probability and time alignment confidence. Stable and clear near-speaking voice is assigned to low-latency transmission, voice segments that are reliable but affected by impacts or strong noise are assigned to redundant protection transmission, and low-reliability segments with insufficient sensor alignment are assigned to delayed buffer transmission. The frame-level control table records the frame sequence number, transmission control category, redundancy flag, mask stability, and alignment confidence, so that the subsequent intercom control processor can read the enhancement and transmission control basis in the same table entry. The advantage of this embodiment is that it binds the voice reliability mask and the link protection level to the same frame-level control table, so that voice preservation, noise suppression, and transmission protection operate based on the same situational result, avoiding the use of different judgment criteria for enhancement processing and link control.

[0023] In one embodiment, the sensor state fusion processor performs reliability encoding on the sensor state vector. The reliability encoding includes sensor sampling interval offset, consecutive missing measurement length, correlation between adjacent sensor features, and synchronization of acoustic feature mutations. When there is a time offset between any sensor feature of the smart sensor and a speech frame, the sensor state fusion processor calculates compensation weights based on the inter-frame change sub-vector, and writes the compensated sensor features, compensation weights, and reliability encoding into the sensor state sub-vector, which is then provided to the state classification processor as an input field for cross-modal consistency determination. Specifically, the sensor sampling interval offset is obtained from the difference between the time interval between two adjacent sensor samplings and the time interval between speech frames, used to characterize the deviation between the sensor sampling rhythm and the speech framing rhythm. The consecutive missing measurement length is obtained from the number of consecutive frames without usable sensor data, used to characterize the sensor state gap. The correlation between adjacent sensor features is obtained by comparing the changing trends of motion intensity, posture change rate, wearing stability, and ambient sound pressure fluctuations in adjacent frames, used for... To characterize the continuity of the sensing state, the synchronization degree of acoustic feature abrupt changes is obtained by comparing the occurrence times of short-term energy abrupt changes, noise spectrum changes, and sensing differences, and is used to characterize whether audio anomalies and sensing anomalies originate from the same source. When there is an offset between the sensing features and the speech frame, the sensing state fusion processor selects adjacent sensing sampling points in the offset direction and compensates for motion intensity, posture change rate, wearing stability, and ambient sound pressure fluctuations, respectively. The compensation weight is determined jointly based on the inter-frame change sub-vector and the time offset. Sensing features with gradual changes are compensated with higher confidence, while sensing features with abrupt changes are compensated with lower confidence. The compensated sensing features are retained side by side with the original sensing features. The situation classification processor reads the compensation weight when making cross-modal consistency judgments to decide whether to use the compensated value, the original value, or reduce the sensing branch participation. The advantage of this embodiment is that by explicitly writing the sensing data quality into the sensing state sub-vector through reliability encoding and compensation weights, short-term sensor delays, missing measurements, or asynchrony will not be mistaken for changes in the real scene, and traceable input is provided for subsequent cross-frame hysteresis and mask smoothing.

[0024] In one embodiment, the cross-modal consistency determination layer is configured with a state constraint matrix consisting of acoustic state labels, sensor cause labels, and wearing stability labels. The state constraint matrix defines the combinatorial relationship between the candidate speech state and the interference cause state. When the candidate speech state contains high-energy abrupt changes and the sensor cause state contains high posture changes or wearing instability, the cross-modal consistency determination layer will output candidates limited to frictional occlusion or short-term impact. When the candidate speech state contains stable harmonics and the wearing stability label satisfies the continuous stability condition, the output candidates will be limited to static proximity or mobile proximity. Specifically, the acoustic state labels are output from the acoustic state branch and may include labels such as stable harmonics, energy abrupt changes, spectral structure disorder, clear speech boundaries, and unclear speech boundaries. The sensor cause labels are output from the sensor cause branch and may include labels such as stable wearing, posture abrupt changes, continuous movement, sound pressure abrupt changes, and wearing instability. The wearing stability label is formed by the continuous states of wearing stability within multiple adjacent frames. The state constraint matrix represents acoustic state labels by rows and transmits signals by columns. The state constraint matrix uses a causal labeling system. Each matrix element records allowed, restricted, or prohibited combinations. Allowed combinations are ranked in the normal scoring process, restricted combinations are ranked only when both time alignment confidence and speech endpoint candidate probability are met, and prohibited combinations are not ranked. When high-energy abrupt changes and high-posture changes occur simultaneously, candidate outputs are restricted to frictional occlusion or short-term impacts to prevent them from being judged as valid near-talk speech solely due to energy increases. When stable harmonics, stable wearing conditions, and continuous speech boundaries occur simultaneously, candidate outputs are restricted to static or moving near-talk speech to prevent far-field sounds under strong noise from entering the near-talk channel through a single voiceprint or energy condition. When ambient sound pressure changes abruptly but the wearing state remains continuous and stable, the state constraint matrix allows noisy near-talk speech as a candidate, allowing the system to retain the possibility of near-talk speech in a noisy background. The advantage of this embodiment is that the state constraint matrix fixes the combinatorial relationship between acoustic state and causal factors into computable rules, reducing abnormal jumps caused by relying entirely on the free output of the classification model, and providing clear cross-modal constraints on the boundaries between frictional occlusion, short-term impacts, and near-talk speech.

[0025] In one embodiment, reference Figure 4The intercom control processor includes an enhancement parameter generation path. This path reads the time-frequency mask matrix, the intercom processing state, and the inter-frame change sub-vector, and generates a noise spectrum update gating, impulse noise suppression coefficient, voice boundary preservation coefficient, and echo cancellation step size for each voice frame. Specifically, the noise spectrum update gating determines the update open interval based on the duration of the interference cause state. The voice boundary preservation coefficient is determined based on the joint distribution of the voice endpoint candidate probability and the wearing state stability, and shares a frame number with the transmission control category. Specifically, the enhancement parameter generation path reads the time-frequency mask matrix and intercom processing state from the frame-level control table and establishes enhancement parameter records using the frame number as the primary key. The noise spectrum update gating determines whether the current voice frame allows updating the background noise estimate. When the interference cause state is a short-term impulse, the update open interval is closed to avoid writing impulse noise into the long-term noise spectrum. When the interference cause state is a continuous strong impulse, the update open interval is closed. When the noise level and voice endpoint candidate probability are low, the update open interval is opened, allowing the background noise spectrum to follow changes in ambient sound pressure. The impulse noise suppression coefficient is jointly determined by the short-time impulse probability, the difference in attitude change rate, and the difference in ambient sound pressure, and is used to suppress burst broadband energy within the frame. The voice boundary preservation coefficient is jointly determined by the voice endpoint candidate probability and the stability of the wearing state. When the voice endpoint candidate probability increases and the wearing state is continuously stable, the voice boundary preservation coefficient remains in the preservation direction to avoid the accidental deletion of consonant boundaries and initial short sounds. The echo cancellation step size is associated with the device interaction state and the transmission control category. When the local playback, intercom trigger, or buffer playback state changes, the echo cancellation step size is updated synchronously with the frame-level control table. The advantage of this embodiment is that the enhanced parameter generation path converts the sensing cause state from the classification result into specific voice processing parameters, so that the noise spectrum update, impulse suppression, voice boundary preservation, and echo cancellation are consistent under the same frame number, reducing misprocessing between motion interference and voice endpoints.

[0026] In one embodiment, the intercom control processor includes a link parameter generation path. This path takes the transmission control category as input, constructs a link parameter group using the coding rate, the forward error correction ratio, the number of key voice frame repetitions, and the receiver buffer depth, and sets a frame sequence number and a protection identifier for the link parameter group. For redundant protection transmission, the link parameter generation path sets the forward error correction ratio based on the length of the continuous reliable region in the time-frequency mask matrix and the difference component of the ambient sound pressure fluctuation, and sets the number of key voice frame repetitions based on the voice endpoint candidate probability. Specifically, the link parameter generation path reads the transmission control category from the frame-level control table. Under the low-latency transmission category, it generates a link parameter group with a shorter receiver buffer depth and a matching coding rate. Under the redundant protection transmission category, it writes the forward error correction ratio and the number of key voice frame repetitions into the protection identifier. Under the delayed buffer transmission category, it merges multiple low-reliability frames into a buffer segment that can be waited for judgment. The link parameter group can be represented as: ; in, This represents the bitrate control amount for the i-th frame. This indicates the forward error correction ratio control value. This indicates the control amount for the number of repetitions of key speech frames. This indicates the receive-end buffer depth control value. This represents the link parameter mapping function. Indicates the send control category, This represents the average retention value of a continuous reliable region in the time-frequency mask matrix. This represents the component of ambient sound pressure fluctuation difference. Represents the candidate probability of the voice endpoint; when Redundancy protection transmission Represents continuous and reliable speech segments and When mutations exist, Will Set to higher protection direction and Pointing to the key speech frame repeats, when Select low latency transmission and When continuously stable, Will The shorter buffer direction is set; the key speech frames are jointly determined by the speech endpoint candidate probability and the continuous reliable region of the time-frequency mask, avoiding the selection of repeated frames only according to the energy level; the advantage of this embodiment is that the link parameter generation path converts the speech reliability, on-site sound pressure changes and transmission category into a unified link parameter group, so that encoding, error correction, repeated transmission and receiving buffer have the same decision source, reducing the problem of insufficient protection of key speech segments by fixed transmission parameters in motion and strong noise scenarios.

[0027] In one embodiment, the situation classification processor sets a cross-frame state hysteresis constraint. This constraint takes the current intercom processing state, the intercom processing state of the previous time window, the time alignment confidence, and the reliability encoding as inputs to generate a state transition permission flag. The state transition permission flag prohibits direct conversion from noisy proximity communication to static proximity communication when the time alignment confidence is lower than the confidence interval given by the cross-frame state hysteresis constraint or when the consecutive missing length exceeds the allowed length. It also smoothly synthesizes the speech confidence mask within adjacent time windows according to the compensation weight. Specifically, the cross-frame state hysteresis constraint is set at the output of the situation classification processor. Its inputs include the intercom processing state with the highest score in the current frame, the intercom processing state confirmed in the previous time window, the time alignment confidence, the consecutive missing length, the correlation of adjacent sensor features, and the synchronization of acoustic feature mutations. When the current frame suggests conversion from noisy proximity communication to static proximity communication, the cross-frame... State hysteresis constraints require that the time alignment confidence level be within the credible range, the length of continuous missing measurements be within the allowable length, and the correlation between adjacent sensor features be able to prove that the ambient sound pressure fluctuation has subsided; otherwise, the state transition permission flag remains in the prohibited state. When the current frame suggests a transition from friction occlusion to mobile proximity, the cross-frame state hysteresis constraints read the continuous recovery of the wearing state stability and combine it with compensation weights to determine whether to allow the transition. The smooth synthesis of the reliable speech mask uses the current mask, the previous time window mask, and compensation weights as inputs. When the sensor data is reliable, the current mask dominates; when there is a shift or missing measurement in the sensor data, the previous time window mask participates in the synthesis with a higher weight, avoiding abnormal and drastic changes in the mask with a single frame. The advantage of this embodiment is that the cross-frame state hysteresis constraints extend the classification result from single-frame judgment to continuous time judgment, making the processing state transition consistent with the sensor reliability and the continuity of the speech mask, reducing frequent switching of processing states caused by running, turning, or short-term occlusion.

[0028] In one embodiment, the intercom control processor sets up a voice frame priority queue associated with the link parameter group. The voice frame priority queue generates a frame-level transmission sequence number based on the transmission control category, the voice endpoint candidate probability, the length of the continuous reliable region, and the difference component of the ambient sound pressure fluctuation. When the transmission buffer control parameter indicates that the buffer occupancy is close to the upper limit, the voice frame priority queue retains continuous voice segments with redundancy protection transmission categories and the corresponding key voice frame repetition counts, and performs delayed merging encoding on frames with low voice endpoint candidate probabilities and lacking wear-state stability support. Specifically, the voice frame priority queue uses the frame sequence number and protection identifier in the link parameter group as indexes to place low-latency transmission frames, redundancy protection transmission frames, and delayed buffer transmission frames into the same sorting space. During sorting, the voice endpoint candidate probability, continuous reliable region length, ambient sound pressure fluctuation difference component, forward error correction ratio control amount, key voice frame repetition count control amount, and receiver buffer depth control amount are read. The frame-level transmission sequence number can be represented as: ; in, This represents the transmission sequence score of the i-th frame. This indicates that the control category mapping value is being sent. Indicates the candidate probability of the voice endpoint. Indicates the length of the continuous reliable region. This represents the absolute value of the environmental sound pressure fluctuation component. This represents the cost of waiting for the cache. to Represents the sorting mapping coefficients; when Corresponding redundancy protection transmission and When representing continuous speech segments, In the priority transmission direction, when When a low value is selected and the wearing stability cannot support near-sound speech, the frame is written into the delayed merging coding area. When the buffer occupancy is close to the upper limit, the voice frame priority queue retains continuous voice segments with redundant protection transmission categories and the corresponding key voice frame repetition counts. Frames with low voice endpoint candidate probability, low mask stability, and lack of wearing stability support are merged, encoded, and sent later. The receiving end restores the playback order based on the frame sequence number and protection identifier in the link parameter group. The advantage of this embodiment is that the voice frame priority queue incorporates frame-level voice reliability, sound pressure change, protection category, and buffer pressure into the sorting, so that continuous segments containing valid near-sound speech can still be retained when the buffer is congested, and the occupation of the transmission window by low-reliability interference frames is reduced.

Claims

1. A voice intercom system for a law enforcement recorder integrating intelligent sensors, characterized in that: This includes a voice frame processor, a sensor state fusion processor, a situation classification processor, and an intercom control processor; The voice frame processor is configured to frame the intercom voice collected by the law enforcement recorder and generate short-time energy, harmonic structure, speech intelligibility, noise spectrum stability and voice endpoint candidate probability. The sensor state fusion processor is configured to receive motion intensity, posture change rate, wearing status stability, ambient sound pressure fluctuation and device interaction status output by smart sensors, and establish a sensor state vector with the intercom voice frame under a unified time reference. The situation classification processor determines the intercom processing status based on the sensing situation vector and generates a voice trust mask and a link protection level. The intercom control processor jointly configures noise suppression weights, coding rate, forward error correction ratio, and transmission buffer control parameters based on the voice trust mask and the link protection level.

2. The law enforcement recorder voice intercom system integrating intelligent sensors according to claim 1, characterized in that, The sensor state fusion processor uses the start and end times of the intercom voice frame as the main index to resample, delay-compensate, and mark the motion intensity, posture change rate, wearing stability, ambient sound pressure fluctuation, and device interaction state, and writes the acoustic features of each voice frame and the sensor features within the same time window into the same vector position. State difference components and sensor missing measurement indicators are retained between adjacent time windows. The sensing state vector includes acoustic state sub-vector, sensing state sub-vector, time alignment confidence, and inter-frame change sub-vector.

3. The law enforcement recorder voice intercom system integrating intelligent sensors according to claim 2, characterized in that, The situation classification processor includes an acoustic state branch, a sensing cause branch, and a cross-modal consistency determination layer. The acoustic state branch generates candidate speech states based on the short-time energy, the harmonic structure, the speech intelligibility, and the noise spectrum stability. The sensing cause branch generates the interference cause state based on the motion intensity, the posture change rate, the wearing state stability, and the ambient sound pressure fluctuation; The cross-modal consistency determination layer jointly determines the candidate speech state and the interference cause state, and outputs one of the following intercom processing states: static near speech, moving near speech, strong noise near speech, friction blockage, short-term impact, or far-field interference.

4. The law enforcement recorder voice intercom system integrating intelligent sensors according to claim 3, characterized in that, The situation classification processor configures voice trust mask generation rules and link protection level generation rules for each of the intercom processing states. The speech credibility mask generation rule maps the near-speech probability, sensor causation state, and inter-frame variation sub-vectors into a time-frequency mask matrix. The link protection level generation rule maps the stability of the time-frequency mask matrix, the voice endpoint candidate probability, and the time alignment confidence to one of the following transmission control categories: low-latency transmission, redundancy protection transmission, or delayed buffer transmission, and writes the transmission control category and the corresponding redundancy identifier into the frame-level control table.

5. The law enforcement recorder voice intercom system integrating intelligent sensors according to claim 2, characterized in that, The sensor state fusion processor performs reliability encoding on the sensor situation vector; The reliability coding includes sensor sampling interval offset, consecutive missing measurement length, correlation between adjacent sensor features, and synchronization of acoustic feature abrupt changes. When there is a time offset between any sensing feature of the smart sensor and the voice frame, the sensing state fusion processor calculates the compensation weight based on the inter-frame change sub-vector, and writes the compensated sensing feature, compensation weight and reliability code together into the sensing state sub-vector, and then provides it to the situation classification processor as the input field for cross-modal consistency determination.

6. The law enforcement recorder voice intercom system integrating intelligent sensors according to claim 3, characterized in that, The cross-modal consistency determination layer is configured with a state constraint matrix consisting of acoustic state tags, sensor cause tags, and wearing stability tags. The state constraint matrix defines the combinatorial relationship between the candidate speech state and the interference cause state; When the candidate speech state contains a high-energy mutation and the sensing cause state contains a high posture change or wearing instability, the cross-modal consistency determination layer will output candidates limited to friction occlusion or short-term impact. When the candidate speech state contains stable harmonics and the wearing of the stable tag satisfies the continuous stability condition, the output candidate will be limited to static proximity or mobile proximity.

7. The law enforcement recorder voice intercom system integrating intelligent sensors according to claim 4, characterized in that, The intercom control processor includes an enhanced parameter generation path; The enhancement parameter generation path reads the time-frequency mask matrix, the intercom processing state, and the inter-frame change sub-vector, and generates noise spectrum update gating, impulse noise suppression coefficient, speech boundary preservation coefficient, and echo cancellation step size for each speech frame; The noise spectrum update gating determines the update open interval based on the duration of the interference cause state, the voice boundary preservation coefficient is determined based on the joint distribution of the voice endpoint candidate probability and the wearing state stability, and shares the frame number with the transmission control category.

8. The law enforcement recorder voice intercom system integrating intelligent sensors according to claim 4, characterized in that, The intercom control processor includes a link parameter generation path; The link parameter generation path takes the transmission control category as input, and constructs a link parameter group by the coding rate, the forward error correction ratio, the number of repetitions of key voice frames and the receiver buffer depth, and sets a frame sequence number and a protection identifier for the link parameter group. For the redundant protection transmission, the link parameter generation path sets the forward error correction ratio based on the length of the continuous reliable region in the time-frequency mask matrix and the differential component of the ambient sound pressure fluctuation, and sets the number of key voice frame repetitions based on the voice endpoint candidate probability.

9. The law enforcement recorder voice intercom system integrating intelligent sensors according to claim 5, characterized in that, The situation classification processor sets cross-frame state hysteresis constraints. The cross-frame state hysteresis constraint takes the current intercom processing state, the intercom processing state of the previous time window, the time alignment confidence, and the reliability code as inputs to generate a state transition permission flag. The state transition permission flag prohibits direct conversion from noisy proximity speech to static proximity speech when the time alignment confidence is lower than the confidence interval given by the cross-frame state hysteresis constraint or the continuous missing length exceeds the allowable length, and the speech confidence mask in adjacent time windows is smoothly synthesized according to the compensation weight.

10. The law enforcement recorder voice intercom system integrating intelligent sensors according to claim 8, characterized in that, The intercom control processor sets up a voice frame priority queue associated with the link parameter group; The voice frame priority queue generates frame-level transmission sequence numbers according to the transmission control category, the voice endpoint candidate probability, the length of the continuous reliable region, and the differential component of the ambient sound pressure fluctuation. When the transmission buffer control parameter indicates that the buffer occupancy is close to the upper limit, the voice frame priority queue retains continuous voice segments with redundancy protection transmission categories and the corresponding key voice frame repetition counts, and performs delayed merging encoding on frames with low voice endpoint candidate probability and lack of wearing state stability support.

Images